Wang Qing [Wed, 9 Feb 2022 08:39:19 +0000 (00:39 -0800)]
net: ethernet: cavium: use div64_u64() instead of do_div()
do_div() does a 64-by-32 division.
When the divisor is u64, do_div() truncates it to 32 bits, this means it
can test non-zero and be truncated to zero for division.
fix do_div.cocci warning:
do_div() does a 64-by-32 division, please consider using div64_u64 instead.
Vladimir Oltean [Wed, 9 Feb 2022 12:04:33 +0000 (14:04 +0200)]
net: dsa: fix panic when DSA master device unbinds on shutdown
Rafael reports that on a system with LX2160A and Marvell DSA switches,
if a reboot occurs while the DSA master (dpaa2-eth) is up, the following
panic can be seen:
It can be seen from the stack trace that the problem is that the
deregistration of the master causes a dev_close(), which gets notified
as NETDEV_GOING_DOWN to dsa_slave_netdevice_event().
But dsa_switch_shutdown() has already run, and this has unregistered the
DSA slave interfaces, and yet, the NETDEV_GOING_DOWN handler attempts to
call dev_close_many() on those slave interfaces, leading to the problem.
The previous attempt to avoid the NETDEV_GOING_DOWN on the master after
dsa_switch_shutdown() was called seems improper. Unregistering the slave
interfaces is unnecessary and unhelpful. Instead, after the slaves have
stopped being uppers of the DSA master, we can now reset to NULL the
master->dsa_ptr pointer, which will make DSA start ignoring all future
notifier events on the master.
Fixes: 0650bf52b31f ("net: dsa: be compatible with masters which unregister on shutdown") Reported-by: Rafael Richter <[email protected]> Signed-off-by: Vladimir Oltean <[email protected]> Signed-off-by: David S. Miller <[email protected]>
David S. Miller [Wed, 9 Feb 2022 13:15:35 +0000 (13:15 +0000)]
Merge branch 'dpaa2-eth-sw-TSO'
Ioana Ciornei says:
====================
dpaa2-eth: add support for software TSO
This series adds support for driver level TSO in the dpaa2-eth driver.
The first 5 patches lay the ground work for the actual feature:
rearrange some variable declaration, cleaning up the interraction with
the S/G Table buffer cache etc.
The 6th patch adds the actual driver level software TSO support by using
the usual tso_build_hdr()/tso_build_data() APIs and creates the S/G FDs.
With this patch set we can see the following improvement in a TCP flow
running on a single [email protected] of the LX2160A SoC:
Ioana Ciornei [Wed, 9 Feb 2022 09:23:35 +0000 (11:23 +0200)]
soc: fsl: dpio: read the consumer index from the cache inhibited area
Once we added support in the dpaa2-eth for driver level software TSO we
observed the following situation: if the EQCR CI (consumer index) is
read from the cache-enabled area we sometimes end up with a computed
value of available enqueue entries bigger than the size of the ring.
This eventually will lead to the multiple enqueue of the same FD which
will determine the same FD to end up on the Tx confirmation path and the
same skb being freed twice.
Just read the consumer index from the cache inhibited area so that we
avoid this situation.
Ioana Ciornei [Wed, 9 Feb 2022 09:23:34 +0000 (11:23 +0200)]
dpaa2-eth: add support for software TSO
This patch adds support for driver level TSO in the enetc driver using
the TSO API.
There is not much to say about this specific implementation. We are
using the usual tso_build_hdr(), tso_build_data() to create each data
segment, we create an array of S/G FDs where the first S/G entry is
referencing the header data and the remaining ones the data portion.
For the S/G Table buffer we use the same cache of buffers used on the
other non-GSO cases - dpaa2_eth_sgt_get() and dpaa2_eth_sgt_recycle().
We cannot keep a DMA coherent buffer for all the TSO headers because the
DPAA2 architecture does not work in a ring based fashion so we just
allocate a buffer each time.
Even with these limitations we get the following improvement in TCP
termination on the LX2160A SoC, on a single A72 core running at 2.2GHz.
Ioana Ciornei [Wed, 9 Feb 2022 09:23:33 +0000 (11:23 +0200)]
dpaa2-eth: work with an array of FDs
Up until now, the __dpaa2_eth_tx function used a single FD on the stack
to construct the structure to be enqueued. Since we are now preparing
the ground work to add support for TSO done in software at the driver
level, the same function needs to work with an array of FDs and enqueue
as many as the build_*_fd functions create.
Make the necessary adjustments in order to do this. These include:
keeping an array of FDs in a percpu structure, cleaning up the necessary
FDs before populating it and then, retrying the enqueue process up till
all the generated FDs were enqueued or until we reach the maximum number
retries.
This patch does not change the fact that only a single FD will result
from a __dpaa2_eth_tx call but rather just creates the necessary changes
for the next patch.
Ioana Ciornei [Wed, 9 Feb 2022 09:23:32 +0000 (11:23 +0200)]
dpaa2-eth: use the S/G table cache also for the normal S/G path
Instead of allocating memory for an S/G table each time a nonlinear skb
is processed, and then freeing it on the Tx confirmation path, use the
S/G table cache in order to reuse the memory.
For this to work we have to change the size of the cached buffers so
that it can hold the maximum number of scatterlist entries.
Other than that, each allocate/free call is replaced by a call to the
dpaa2_eth_sgt_get/dpaa2_eth_sgt_recycle functions, introduced in the
previous patch.
Ioana Ciornei [Wed, 9 Feb 2022 09:23:31 +0000 (11:23 +0200)]
dpaa2-eth: extract the S/G table buffer cache interaction into functions
The dpaa2-eth driver uses in certain circumstances a buffer cache for
the S/G tables needed in case of a S/G FD. At the moment, the
interraction with the cache is open-coded and couldn't be reused easily.
Add two new functions - dpaa2_eth_sgt_get and dpaa2_eth_sgt_recycle -
which help with code reusability.
Ioana Ciornei [Wed, 9 Feb 2022 09:23:29 +0000 (11:23 +0200)]
dpaa2-eth: rearrange variable declaration in __dpaa2_eth_tx
In the next patches we'll be moving things arroung in the mentioned
function and also add some new variable declarations. Before all this,
cleanup the variable declaration order.
David S. Miller [Wed, 9 Feb 2022 13:02:33 +0000 (13:02 +0000)]
Merge branch 'octeontx2-af-priority-flow-control'
Hariprasad Kelam says:
====================
Priority flow control support for RVU netdev
In network congestion, instead of pausing all traffic on link
PFC allows user to selectively pause traffic according to its
class. This series of patches add support of PFC for RVU netdev
drivers.
Patch1 adds support to disable pause frames by default as
with PFC user can enable either PFC or 802.3 pause frames.
Patch2&3 adds resource management support for flow control
and configures necessary registers for PFC.
Patch4 adds dcb ops registration for netdev drivers.
Data centric bridging designed to eliminate packet loss due to
queue overflow by adding enhancements to ethernet network such as
proprity flow control etc. This patch adds support for management
of Priority flow control(PFC) on Octeontx2 and CN10K interfaces.
To enable PFC for all priorities
dcb pfc set dev eth0 prio-pfc all:on/off
To enable PFC on selected priorites
dcb pfc set dev eth0 prio-pfc 0:on/off 1:on/off ..7:on/off
With the ntuple commands user can map Priority to receive queues.
On queue overflow NIX will assert backpressure such that PFC pause frames
are genarated with mapped priority.
To map priority 7 to Queue 1
ethtool -U eth0 flow-type ether dst xx:xx:xx:xx:xx:xx vlan 0xe00a
m 0x1fff queue 1
CN10K MAC block (RPM) and Octeontx2 MAC block (CGX) both supports
PFC flow control and 802.3X flow control pause frames.
Each MAC block supports max 4 LMACS and AF driver assigns same
(MAC,LMAC) to PF and its VFs. As PF and its share same (MAC,LMAC)
pair we need resource management to address below scenarios
1. Maintain PFC and 8023X pause frames mutually exclusive.
2. Reject disable flow control request if other PF or Vfs
enabled it.
octeontx2-af: Priority flow control configuration support
Prirority based flow control (802.1Qbb) mechanism is similar to
ethernet pause frames (802.3x) instead pausing all traffic on a link,
PFC allows user to selectively pause traffic according to its class.
Oceteontx2 MAC block (CGX) and CN10K Mac block (RPM) both supports
PFC. As upper layer mbox handler is same for both the MACs, this
patch configures PFC by calling apporopritate callbacks.
octeontx2-af: Don't enable Pause frames by default
Current implementation is such that 802.3x pause frames are
enabled by default. As CGX and RPM blocks support PFC
(priority flow control) also, instead of driver enabling one
between them enable them upon request from PF or its VFs.
Also add support to disable pause frames in driver unbind.
Originally we proposed a new hdmi-5v-supply regulator reference
for CI20 device tree but that was superseded by a better idea to use
the already defined "ddc-en-gpios" property of the "hdmi-connector".
Since "MIPS: DTS: CI20: Add DT nodes for HDMI setup" has already
been applied to v5.17-rc1, we add this on top.
Fixes: ae1b8d2c2de9 ("MIPS: DTS: CI20: Add DT nodes for HDMI setup") Signed-off-by: H. Nikolaus Schaller <[email protected]> Reviewed-by: Paul Cercueil <[email protected]> Signed-off-by: Thomas Bogendoerfer <[email protected]>
Raju Rangoju [Wed, 9 Feb 2022 04:32:01 +0000 (10:02 +0530)]
net: amd-xgbe: disable interrupts during pci removal
Hardware interrupts are enabled during the pci probe, however,
they are not disabled during pci removal.
Disable all hardware interrupts during pci removal to avoid any
issues.
Fixes: e75377404726 ("amd-xgbe: Update PCI support to use new IRQ functions") Suggested-by: Selwin Sebastian <[email protected]> Signed-off-by: Raju Rangoju <[email protected]> Signed-off-by: David S. Miller <[email protected]>
Jon Maloy [Wed, 9 Feb 2022 03:22:37 +0000 (22:22 -0500)]
tipc: rate limit warning for received illegal binding update
It would be easy to craft a message containing an illegal binding table
update operation. This is handled correctly by the code, but the
corresponding warning printout is not rate limited as is should be.
We fix this now.
Fixes: b97bf3fd8f6a ("[TIPC] Initial merge") Signed-off-by: Jon Maloy <[email protected]> Signed-off-by: David S. Miller <[email protected]>
Reported by Kernel Concurrency Sanitizer on:
CPU: 1 PID: 23563 Comm: syz-executor.5 Not tainted 5.17.0-rc2-syzkaller-00064-gc36c04c2e132 #0
Hardware name: Google Google Compute Engine/Google Compute Engine, BIOS Google 01/01/2011
David S. Miller [Wed, 9 Feb 2022 12:00:11 +0000 (12:00 +0000)]
Merge branch 'MCTP-tag-control-interface'
Jeremy Kerr says:
====================
MCTP tag control interface
This series implements a small interface for userspace-controlled
message tag allocation for the MCTP protocol. Rather than leaving the
kernel to allocate per-message tag values, userspace can explicitly
allocate (and release) message tags through two new ioctls:
SIOCMCTPALLOCTAG and SIOCMCTPDROPTAG.
In order to do this, we first introduce some minor changes to the tag
handling, including a couple of new tests for the route input paths.
As always, any comments/queries/etc are most welcome.
v2:
- make mctp_lookup_prealloc_tag static
- minor checkpatch formatting fixes
====================
Matt Johnston [Wed, 9 Feb 2022 04:05:57 +0000 (12:05 +0800)]
mctp: Add SIOCMCTP{ALLOC,DROP}TAG ioctls for tag control
This change adds a couple of new ioctls for mctp sockets:
SIOCMCTPALLOCTAG and SIOCMCTPDROPTAG. These ioctls provide facilities
for explicit allocation / release of tags, overriding the automatic
allocate-on-send/release-on-reply and timeout behaviours. This allows
userspace more control over messages that may not fit a simple
request/response model.
In order to indicate a pre-allocated tag to the sendmsg() syscall, we
introduce a new flag to the struct sockaddr_mctp.smctp_tag value:
MCTP_TAG_PREALLOC.
Jeremy Kerr [Wed, 9 Feb 2022 04:05:54 +0000 (12:05 +0800)]
mctp: tests: Add key state tests
This change adds a few more tests to check the key/tag lookups on route
input. We add a specific entry to the keys lists, route a packet with
specific header values, and check for key match/mismatch.
Jeremy Kerr [Wed, 9 Feb 2022 04:05:53 +0000 (12:05 +0800)]
mctp: tests: Rename FL_T macro to FL_TO
This is a definition for the tag-owner flag, which has TO as a standard
abbreviation. We'll want to add a helper for the actual tag value in a
future change.
David S. Miller [Wed, 9 Feb 2022 11:57:54 +0000 (11:57 +0000)]
Merge branch '40GbE' of git://git.kernel.org/pub/scm/linux/kernel/git/tnguy/next
-queue
Tony Nguyen says:
====================
40GbE Intel Wired LAN Driver Updates 2022-02-08
Joe Damato says:
This patch set makes several updates to the i40e driver stats collection
and reporting code to help users of i40e get a better sense of how the
driver is performing and interacting with the rest of the kernel.
These patches include some new stats (like waived and busy) which were
inspired by other drivers that track stats using the same nomenclature.
The new stats and an existing stat, rx_reuse, are now accessible with
ethtool to make harvesting this data more convenient for users.
====================
Duoming Zhou [Tue, 8 Feb 2022 15:40:00 +0000 (23:40 +0800)]
ax25: fix NPD bug in ax25_disconnect
The ax25_disconnect() in ax25_kill_by_device() is not
protected by any locks, thus there is a race condition
between ax25_disconnect() and ax25_destroy_socket().
when ax25->sk is assigned as NULL by ax25_destroy_socket(),
a NULL pointer dereference bug will occur if site (1) or (2)
dereferences ax25->sk.
Tianyu Lan [Tue, 8 Feb 2022 14:26:52 +0000 (09:26 -0500)]
Netvsc: Call hv_unmap_memory() in the netvsc_device_remove()
netvsc_device_remove() calls vunmap() inside which should not be
called in the interrupt context. Current code calls hv_unmap_memory()
in the free_netvsc_device() which is rcu callback and maybe called
in the interrupt context. This will trigger BUG_ON(in_interrupt())
in the vunmap(). Fix it via moving hv_unmap_memory() to netvsc_device_
remove().
Fixes: 846da38de0e8 ("net: netvsc: Add Isolation VM support for netvsc driver") Signed-off-by: Tianyu Lan <[email protected]> Signed-off-by: David S. Miller <[email protected]>
David S. Miller [Wed, 9 Feb 2022 11:41:48 +0000 (11:41 +0000)]
Merge branch 'net-fix-skb-unclone-issues'
Antoine Tenart says:
====================
net: fix issues when uncloning an skb dst+metadata
This fixes two issues when uncloning an skb dst+metadata in
tun_dst_unclone; this was initially reported by Vlad Buslov[1]. Because
of the memory leak fixed by patch 2, the issue in patch 1 never happened
in practice.
tun_dst_unclone is called from two different places, one in geneve/vxlan
to handle PMTU and one in net/openvswitch/actions.c where it is used to
retrieve tunnel information. While both Vlad and I tested the former, we
could not for the latter. I did spend quite some time trying to, but
that code path is not easy to trigger. Code inspection shows this should
be fine, the tunnel information (dst+metadata) is uncloned and the skb
it is referenced from is only consumed after all accesses to the tunnel
information are done:
do_execute_actions
output_userspace
dev_fill_metadata_dst <- dst+metadata is uncloned
ovs_dp_upcall
queue_userspace_packet
ovs_nla_put_tunnel_info <- metadata (tunnel info) is accessed
consume_skb <- dst+metadata is freed
Antoine Tenart [Mon, 7 Feb 2022 17:13:19 +0000 (18:13 +0100)]
net: fix a memleak when uncloning an skb dst and its metadata
When uncloning an skb dst and its associated metadata, a new
dst+metadata is allocated and later replaces the old one in the skb.
This is helpful to have a non-shared dst+metadata attached to a specific
skb.
The issue is the uncloned dst+metadata is initialized with a refcount of
1, which is increased to 2 before attaching it to the skb. When
tun_dst_unclone returns, the dst+metadata is only referenced from a
single place (the skb) while its refcount is 2. Its refcount will never
drop to 0 (when the skb is consumed), leading to a memory leak.
Fix this by removing the call to dst_hold in tun_dst_unclone, as the
dst+metadata refcount is already 1.
Antoine Tenart [Mon, 7 Feb 2022 17:13:18 +0000 (18:13 +0100)]
net: do not keep the dst cache when uncloning an skb dst and its metadata
When uncloning an skb dst and its associated metadata a new dst+metadata
is allocated and the tunnel information from the old metadata is copied
over there.
The issue is the tunnel metadata has references to cached dst, which are
copied along the way. When a dst+metadata refcount drops to 0 the
metadata is freed including the cached dst entries. As they are also
referenced in the initial dst+metadata, this ends up in UaFs.
In practice the above did not happen because of another issue, the
dst+metadata was never freed because its refcount never dropped to 0
(this will be fixed in a subsequent patch).
Fix this by initializing the dst cache after copying the tunnel
information from the old metadata to also unshare the dst cache.
Document Gigabit Ethernet IP found on RZ/G2UL SoC. Gigabit Ethernet
Interface is identical to one found on the RZ/G2L SoC. No driver changes
are required as generic compatible string "renesas,rzg2l-gbeth" will be
used as a fallback.
Document Gigabit Ethernet IP found on RZ/V2L SoC. Gigabit Ethernet
Interface is identical to one found on the RZ/G2L SoC. No driver changes
are required as generic compatible string "renesas,rzg2l-gbeth" will be
used as a fallback.
netfilter: nft_cmp: optimize comparison for 16-bytes
Allow up to 16-byte comparisons with a new cmp fast version. Use two
64-bit words and calculate the mask representing the bits to be
compared. Make sure the comparison is 64-bit aligned and avoid
out-of-bound memory access on registers.
Oliver Hartkopp [Wed, 9 Feb 2022 07:36:01 +0000 (08:36 +0100)]
can: isotp: fix error path in isotp_sendmsg() to unlock wait queue
Commit 43a08c3bdac4 ("can: isotp: isotp_sendmsg(): fix TX buffer concurrent
access in isotp_sendmsg()") introduced a new locking scheme that may render
the userspace application in a locking state when an error is detected.
This issue shows up under high load on simultaneously running isotp channels
with identical configuration which is against the ISO specification and
therefore breaks any reasonable PDU communication anyway.
Oliver Hartkopp [Tue, 8 Feb 2022 20:00:26 +0000 (21:00 +0100)]
can: isotp: fix potential CAN frame reception race in isotp_rcv()
When receiving a CAN frame the current code logic does not consider
concurrently receiving processes which do not show up in real world
usage.
Ziyang Xuan writes:
The following syz problem is one of the scenarios. so->rx.len is
changed by isotp_rcv_ff() during isotp_rcv_cf(), so->rx.len equals
0 before alloc_skb() and equals 4096 after alloc_skb(). That will
trigger skb_over_panic() in skb_put().
Therefore we make sure the state changes and data structures stay
consistent at CAN frame reception time by adding a spin_lock in
isotp_rcv(). This fixes the issue reported by syzkaller but does not
affect real world operation.
Hengqi Chen [Mon, 7 Feb 2022 14:31:33 +0000 (22:31 +0800)]
libbpf: Add BPF_KPROBE_SYSCALL macro
Add syscall-specific variant of BPF_KPROBE named BPF_KPROBE_SYSCALL ([0]).
The new macro hides the underlying way of getting syscall input arguments.
With the new macro, the following code:
SEC("kprobe/__x64_sys_close")
int BPF_KPROBE(do_sys_close, struct pt_regs *regs)
{
int fd;
fd = PT_REGS_PARM1_CORE(regs);
/* do something with fd */
}
can be written as:
SEC("kprobe/__x64_sys_close")
int BPF_KPROBE_SYSCALL(do_sys_close, int fd)
{
/* do something with fd */
}
Andrii Nakryiko [Wed, 9 Feb 2022 05:16:15 +0000 (21:16 -0800)]
Merge branch 'Fix accessing syscall arguments'
Ilya Leoshkevich says:
====================
libbpf now has macros to access syscall arguments in an
architecture-agnostic manner, but unfortunately they have a number of
issues on non-Intel arches, which this series aims to fix.
v1: https://lore.kernel.org/bpf/20220201234200.1836443[email protected]/
v1 -> v2:
* Put orig_gpr2 in place of args[1] on s390 (Vasily).
* Fix arm64, powerpc and riscv (Heiko).
v3: https://lore.kernel.org/bpf/20220204145018.1983773[email protected]/
v3 -> v4:
* Undo arm64's and s390's user_pt_regs changes.
* Use struct pt_regs when vmlinux.h is available (Andrii).
* Use offsetofend for accessing orig_gpr2 and orig_x0 (Andrii).
* Move libbpf's copy of offsetofend to a new header.
* Fix riscv's __PT_FP_REG.
* Use PT_REGS_SYSCALL_REGS in test_probe_user.c.
* Test bpf_syscall_macro with userspace headers.
* Use Naveen's suggestions and code in patches 5 and 6.
* Add warnings to arm64's and s390's ptrace.h (Andrii).
v4: https://lore.kernel.org/bpf/20220208051635.2160304[email protected]/
v4 -> v5:
* Go back to v3.
* Do not touch arch headers.
* Use CO-RE struct flavors to access orig_x0 and orig_gpr2.
* Fail compilation if non-CO-RE macros are used to access the first
syscall parameter on arm64 and s390.
* Fix accessing frame pointer on riscv.
====================
libbpf: Fix accessing the first syscall argument on s390
On s390, the first syscall argument should be accessed via orig_gpr2
(see arch/s390/include/asm/syscall.h). Currently gpr[2] is used
instead, leading to bpf_syscall_macro test failure.
orig_gpr2 cannot be added to user_pt_regs, since its layout is a part
of the ABI. Therefore provide access to it only through
PT_REGS_PARM1_CORE_SYSCALL() by using a struct pt_regs flavor.
libbpf: Fix accessing the first syscall argument on arm64
On arm64, the first syscall argument should be accessed via orig_x0
(see arch/arm64/include/asm/syscall.h). Currently regs[0] is used
instead, leading to bpf_syscall_macro test failure.
orig_x0 cannot be added to struct user_pt_regs, since its layout is a
part of the ABI. Therefore provide access to it only through
PT_REGS_PARM1_CORE_SYSCALL() by using a struct pt_regs flavor.
riscv does not select ARCH_HAS_SYSCALL_WRAPPER, so its syscall
handlers take "unpacked" syscall arguments. Indicate this to libbpf
using PT_REGS_SYSCALL_REGS macro.
riscv registers are accessed via struct user_regs_struct, not struct
pt_regs. The program counter member in this struct is called pc, not
epc. The frame pointer is called s0, not fp.
libbpf: Fix accessing syscall arguments on powerpc
powerpc does not select ARCH_HAS_SYSCALL_WRAPPER, so its syscall
handlers take "unpacked" syscall arguments. Indicate this to libbpf
using PT_REGS_SYSCALL_REGS macro.
Architectures that select ARCH_HAS_SYSCALL_WRAPPER pass a pointer to
struct pt_regs to syscall handlers, others unpack it into individual
function parameters. Introduce a macro to describe what a particular
arch does.
selftests/bpf: Fix an endianness issue in bpf_syscall_macro test
bpf_syscall_macro reads a long argument into an int variable, which
produces a wrong value on big-endian systems. Fix by reading the
argument into an intermediate long variable first.
Louis Peens [Tue, 8 Feb 2022 10:14:53 +0000 (11:14 +0100)]
nfp: flower: fix ida_idx not being released
When looking for a global mac index the extra NFP_TUN_PRE_TUN_IDX_BIT
that gets set if nfp_flower_is_supported_bridge is true is not taken
into account. Consequently the path that should release the ida_index
in cleanup is never triggered, causing messages like:
nfp 0000:02:00.0: nfp: Failed to offload MAC on br-ex.
nfp 0000:02:00.0: nfp: Failed to offload MAC on br-ex.
nfp 0000:02:00.0: nfp: Failed to offload MAC on br-ex.
after NFP_MAX_MAC_INDEX number of reconfigs. Ultimately this lead to
new tunnel flows not being offloaded.
Fix this by unsetting the NFP_TUN_PRE_TUN_IDX_BIT before checking if
the port is of type OTHER.
In this series, I made network namespace deletions more scalable,
by 4x on the little benchmark described in this cover letter.
- Remove bottleneck on ipv6 addrconf, by replacing a global
hash table to a per netns one.
- Rework many (struct pernet_operations)->exit() handlers to
exit_batch() ones. This removes many rtnl acquisitions,
and gives to cleanup_net() kind of a priority over rtnl
ownership.
Tested on a host with 24 cpus (48 HT)
Test script:
for nr in {1..10}
do
(for i in {1..10000}; do unshare -n /bin/bash -c "ifconfig lo up"; done) &
done
wait
for i in {1..10}
do
sleep 1
echo 3 >/proc/sys/vm/drop_caches
grep net_namespace /proc/slabinfo
done
Before: We can see host struggles to clean the netns, even after there are no new creations.
Memory cost is high, because each netns consumes a good amount of memory.
Eric Dumazet [Tue, 8 Feb 2022 04:50:37 +0000 (20:50 -0800)]
bonding: switch bond_net_exit() to batch mode
cleanup_net() is competing with other rtnl users.
Batching bond_net_exit() factorizes all rtnl acquistions
to a single one, giving chance for cleanup_net()
to progress much faster, holding rtnl a bit longer.
Eric Dumazet [Tue, 8 Feb 2022 04:50:36 +0000 (20:50 -0800)]
can: gw: switch cangw_pernet_exit() to batch mode
cleanup_net() is competing with other rtnl users.
Avoiding to acquire rtnl for each netns before calling
cgw_remove_all_jobs() gives chance for cleanup_net()
to progress much faster, holding rtnl a bit longer.
Eric Dumazet [Tue, 8 Feb 2022 04:50:35 +0000 (20:50 -0800)]
ipmr: introduce ipmr_net_exit_batch()
cleanup_net() is competing with other rtnl users.
Avoiding to acquire rtnl for each netns before calling
ipmr_rules_exit() gives chance for cleanup_net()
to progress much faster, holding rtnl a bit longer.
Eric Dumazet [Tue, 8 Feb 2022 04:50:34 +0000 (20:50 -0800)]
ip6mr: introduce ip6mr_net_exit_batch()
cleanup_net() is competing with other rtnl users.
Avoiding to acquire rtnl for each netns before calling
ip6mr_rules_exit() gives chance for cleanup_net()
to progress much faster, holding rtnl a bit longer.
Eric Dumazet [Tue, 8 Feb 2022 04:50:33 +0000 (20:50 -0800)]
ipv6: change fib6_rules_net_exit() to batch mode
cleanup_net() is competing with other rtnl users.
fib6_rules_net_exit() seems a good candidate for exit_batch(),
as this gives chance for cleanup_net() to progress much faster,
holding rtnl a bit longer.
Eric Dumazet [Tue, 8 Feb 2022 04:50:29 +0000 (20:50 -0800)]
ipv6/addrconf: use one delayed work per netns
Next step for using per netns inet6_addr_lst
is to have per netns work item to ultimately
call addrconf_verify_rtnl() and addrconf_verify()
with a new 'struct net*' argument.
Everything is still using the global inet6_addr_lst[] table.
If __ibmvnic_open() encounters an error such as when setting link state,
it calls release_resources() which frees the napi structures needlessly.
Instead, have __ibmvnic_open() only clean up the work it did so far (i.e.
disable napi and irqs) and leave the rest to the callers.
If caller of __ibmvnic_open() is ibmvnic_open(), it should release the
resources immediately. If the caller is do_reset() or do_hard_reset(),
they will release the resources on the next reset.
This fixes following crash that occurred when running the drmgr command
several times to add/remove a vnic interface:
====================
More DSA fixes for devres + mdiobus_{alloc,register}
The initial patch series "[net,0/2] Fix mdiobus users with devres"
https://patchwork.kernel.org/project/netdevbpf/cover/20210920214209.1733768[email protected]/
fixed some instances where DSA drivers on slow buses (SPI, I2C) trigger
a panic (changed since then to a warn) in mdiobus_free. That was due to
devres calling mdiobus_free() with no prior mdiobus_unregister(), which
again was due to commit ac3a68d56651 ("net: phy: don't abuse devres in
devm_mdiobus_register()") by Bartosz Golaszewski.
Rafael Richter and Daniel Klauer report yet another variation on that
theme, but this time it applies to any DSA switch driver, not just those
on buses which have a "->shutdown() calls ->remove() which unregisters
children" sequence.
Their setup is that of an LX2160A DPAA2 SoC driving a Marvell DSA switch
(MDIO). DPAA2 Ethernet drivers probe on the "fsl-mc" bus
(drivers/bus/fsl-mc/fsl-mc-bus.c). This bus is meant to be the
kernel-side representation of the networking objects kept by the
Management Complex (MC) firmware.
which proceeds to remove the children on the bus. Among those children,
the dpaa2-eth network driver.
When dpaa2-eth is a DSA master, this removal of the master on shutdown
trips up the device link created by dsa_master_setup(), and as such, the
Marvell switch is also removed.
From this point on, readers can revisit the description of commits 74b6d7d13307 ("net: dsa: realtek: register the MDIO bus under devres") 5135e96a3dd2 ("net: dsa: don't allocate the slave_mii_bus using devres")
since the prerequisites for the BUG_ON in mdiobus_free() have been
accomplished if there is a devres mismatch between mdiobus_alloc() and
mdiobus_register().
Most DSA drivers have this kind of mismatch, and upon my initial
assessment I had not realized the possibility described above, so I
didn't fix it. This patch series walks through all drivers and makes
them use either fully devres, or no devres.
I am aware that there are DSA drivers that are only known to be tested
with a single DSA master, so some patches are probably overkill for
them. But code is copy-pasted from so many sources without fully
understanding the differences, that I think it's better to not leave an
in-tree source of inspiration that may lead to subtle breakage if not
adapted properly.
====================
Vladimir Oltean [Mon, 7 Feb 2022 16:15:53 +0000 (18:15 +0200)]
net: dsa: lantiq_gswip: don't use devres for mdiobus
As explained in commits: 74b6d7d13307 ("net: dsa: realtek: register the MDIO bus under devres") 5135e96a3dd2 ("net: dsa: don't allocate the slave_mii_bus using devres")
mdiobus_free() will panic when called from devm_mdiobus_free() <-
devres_release_all() <- __device_release_driver(), and that mdiobus was
not previously unregistered.
The GSWIP switch is a platform device, so the initial set of constraints
that I thought would cause this (I2C or SPI buses which call ->remove on
->shutdown) do not apply. But there is one more which applies here.
If the DSA master itself is on a bus that calls ->remove from ->shutdown
(like dpaa2-eth, which is on the fsl-mc bus), there is a device link
between the switch and the DSA master, and device_links_unbind_consumers()
will unbind the GSWIP switch driver on shutdown.
So the same treatment must be applied to all DSA switch drivers, which
is: either use devres for both the mdiobus allocation and registration,
or don't use devres at all.
The gswip driver has the code structure in place for orderly mdiobus
removal, so just replace devm_mdiobus_alloc() with the non-devres
variant, and add manual free where necessary, to ensure that we don't
let devres free a still-registered bus.
Fixes: ac3a68d56651 ("net: phy: don't abuse devres in devm_mdiobus_register()") Signed-off-by: Vladimir Oltean <[email protected]> Reviewed-by: Florian Fainelli <[email protected]> Signed-off-by: Jakub Kicinski <[email protected]>
Vladimir Oltean [Mon, 7 Feb 2022 16:15:52 +0000 (18:15 +0200)]
net: dsa: mt7530: fix kernel bug in mdiobus_free() when unbinding
Nobody in this driver calls mdiobus_unregister(), which is necessary if
mdiobus_register() completes successfully. So if the devres callbacks
that free the mdiobus get invoked (this is the case when unbinding the
driver), mdiobus_free() will BUG if the mdiobus is still registered,
which it is.
My speculation is that this is due to the fact that prior to commit ac3a68d56651 ("net: phy: don't abuse devres in devm_mdiobus_register()")
from June 2020, _devm_mdiobus_free() used to call mdiobus_unregister().
But at the time that the mt7530 support was introduced in May 2021, the
API was already changed. It's therefore likely that the blamed patch was
developed on an older tree, and incorrectly adapted to net-next. This
makes the Fixes: tag correct.
Fix the problem by using the devres variant of mdiobus_register.
Vladimir Oltean [Mon, 7 Feb 2022 16:15:51 +0000 (18:15 +0200)]
net: dsa: seville: register the mdiobus under devres
As explained in commits: 74b6d7d13307 ("net: dsa: realtek: register the MDIO bus under devres") 5135e96a3dd2 ("net: dsa: don't allocate the slave_mii_bus using devres")
mdiobus_free() will panic when called from devm_mdiobus_free() <-
devres_release_all() <- __device_release_driver(), and that mdiobus was
not previously unregistered.
The Seville VSC9959 switch is a platform device, so the initial set of
constraints that I thought would cause this (I2C or SPI buses which call
->remove on ->shutdown) do not apply. But there is one more which
applies here.
If the DSA master itself is on a bus that calls ->remove from ->shutdown
(like dpaa2-eth, which is on the fsl-mc bus), there is a device link
between the switch and the DSA master, and device_links_unbind_consumers()
will unbind the seville switch driver on shutdown.
So the same treatment must be applied to all DSA switch drivers, which
is: either use devres for both the mdiobus allocation and registration,
or don't use devres at all.
The seville driver has a code structure that could accommodate both the
mdiobus_unregister and mdiobus_free calls, but it has an external
dependency upon mscc_miim_setup() from mdio-mscc-miim.c, which calls
devm_mdiobus_alloc_size() on its behalf. So rather than restructuring
that, and exporting yet one more symbol mscc_miim_teardown(), let's work
with devres and replace of_mdiobus_register with the devres variant.
When we use all-devres, we can ensure that devres doesn't free a
still-registered bus (it either runs both callbacks, or none).
Fixes: ac3a68d56651 ("net: phy: don't abuse devres in devm_mdiobus_register()") Signed-off-by: Vladimir Oltean <[email protected]> Reviewed-by: Florian Fainelli <[email protected]> Signed-off-by: Jakub Kicinski <[email protected]>
Vladimir Oltean [Mon, 7 Feb 2022 16:15:50 +0000 (18:15 +0200)]
net: dsa: felix: don't use devres for mdiobus
As explained in commits: 74b6d7d13307 ("net: dsa: realtek: register the MDIO bus under devres") 5135e96a3dd2 ("net: dsa: don't allocate the slave_mii_bus using devres")
mdiobus_free() will panic when called from devm_mdiobus_free() <-
devres_release_all() <- __device_release_driver(), and that mdiobus was
not previously unregistered.
The Felix VSC9959 switch is a PCI device, so the initial set of
constraints that I thought would cause this (I2C or SPI buses which call
->remove on ->shutdown) do not apply. But there is one more which
applies here.
If the DSA master itself is on a bus that calls ->remove from ->shutdown
(like dpaa2-eth, which is on the fsl-mc bus), there is a device link
between the switch and the DSA master, and device_links_unbind_consumers()
will unbind the felix switch driver on shutdown.
So the same treatment must be applied to all DSA switch drivers, which
is: either use devres for both the mdiobus allocation and registration,
or don't use devres at all.
The felix driver has the code structure in place for orderly mdiobus
removal, so just replace devm_mdiobus_alloc_size() with the non-devres
variant, and add manual free where necessary, to ensure that we don't
let devres free a still-registered bus.
Fixes: ac3a68d56651 ("net: phy: don't abuse devres in devm_mdiobus_register()") Signed-off-by: Vladimir Oltean <[email protected]> Reviewed-by: Florian Fainelli <[email protected]> Signed-off-by: Jakub Kicinski <[email protected]>
Vladimir Oltean [Mon, 7 Feb 2022 16:15:49 +0000 (18:15 +0200)]
net: dsa: bcm_sf2: don't use devres for mdiobus
As explained in commits: 74b6d7d13307 ("net: dsa: realtek: register the MDIO bus under devres") 5135e96a3dd2 ("net: dsa: don't allocate the slave_mii_bus using devres")
mdiobus_free() will panic when called from devm_mdiobus_free() <-
devres_release_all() <- __device_release_driver(), and that mdiobus was
not previously unregistered.
The Starfighter 2 is a platform device, so the initial set of
constraints that I thought would cause this (I2C or SPI buses which call
->remove on ->shutdown) do not apply. But there is one more which
applies here.
If the DSA master itself is on a bus that calls ->remove from ->shutdown
(like dpaa2-eth, which is on the fsl-mc bus), there is a device link
between the switch and the DSA master, and device_links_unbind_consumers()
will unbind the bcm_sf2 switch driver on shutdown.
So the same treatment must be applied to all DSA switch drivers, which
is: either use devres for both the mdiobus allocation and registration,
or don't use devres at all.
The bcm_sf2 driver has the code structure in place for orderly mdiobus
removal, so just replace devm_mdiobus_alloc() with the non-devres
variant, and add manual free where necessary, to ensure that we don't
let devres free a still-registered bus.
Fixes: ac3a68d56651 ("net: phy: don't abuse devres in devm_mdiobus_register()") Signed-off-by: Vladimir Oltean <[email protected]> Reviewed-by: Florian Fainelli <[email protected]> Signed-off-by: Jakub Kicinski <[email protected]>
Vladimir Oltean [Mon, 7 Feb 2022 16:15:48 +0000 (18:15 +0200)]
net: dsa: ar9331: register the mdiobus under devres
As explained in commits: 74b6d7d13307 ("net: dsa: realtek: register the MDIO bus under devres") 5135e96a3dd2 ("net: dsa: don't allocate the slave_mii_bus using devres")
mdiobus_free() will panic when called from devm_mdiobus_free() <-
devres_release_all() <- __device_release_driver(), and that mdiobus was
not previously unregistered.
The ar9331 is an MDIO device, so the initial set of constraints that I
thought would cause this (I2C or SPI buses which call ->remove on
->shutdown) do not apply. But there is one more which applies here.
If the DSA master itself is on a bus that calls ->remove from ->shutdown
(like dpaa2-eth, which is on the fsl-mc bus), there is a device link
between the switch and the DSA master, and device_links_unbind_consumers()
will unbind the ar9331 switch driver on shutdown.
So the same treatment must be applied to all DSA switch drivers, which
is: either use devres for both the mdiobus allocation and registration,
or don't use devres at all.
The ar9331 driver doesn't have a complex code structure for mdiobus
removal, so just replace of_mdiobus_register with the devres variant in
order to be all-devres and ensure that we don't free a still-registered
bus.
Fixes: ac3a68d56651 ("net: phy: don't abuse devres in devm_mdiobus_register()") Signed-off-by: Vladimir Oltean <[email protected]> Reviewed-by: Florian Fainelli <[email protected]> Tested-by: Oleksij Rempel <[email protected]> Signed-off-by: Jakub Kicinski <[email protected]>
Vladimir Oltean [Mon, 7 Feb 2022 16:15:47 +0000 (18:15 +0200)]
net: dsa: mv88e6xxx: don't use devres for mdiobus
As explained in commits: 74b6d7d13307 ("net: dsa: realtek: register the MDIO bus under devres") 5135e96a3dd2 ("net: dsa: don't allocate the slave_mii_bus using devres")
mdiobus_free() will panic when called from devm_mdiobus_free() <-
devres_release_all() <- __device_release_driver(), and that mdiobus was
not previously unregistered.
The mv88e6xxx is an MDIO device, so the initial set of constraints that
I thought would cause this (I2C or SPI buses which call ->remove on
->shutdown) do not apply. But there is one more which applies here.
If the DSA master itself is on a bus that calls ->remove from ->shutdown
(like dpaa2-eth, which is on the fsl-mc bus), there is a device link
between the switch and the DSA master, and device_links_unbind_consumers()
will unbind the Marvell switch driver on shutdown.
systemd-shutdown[1]: Powering off.
mv88e6085 0x0000000008b96000:00 sw_gl0: Link is Down
fsl-mc dpbp.9: Removing from iommu group 7
fsl-mc dpbp.8: Removing from iommu group 7
------------[ cut here ]------------
kernel BUG at drivers/net/phy/mdio_bus.c:677!
Internal error: Oops - BUG: 0 [#1] PREEMPT SMP
Modules linked in:
CPU: 0 PID: 1 Comm: systemd-shutdow Not tainted 5.16.5-00040-gdc05f73788e5 #15
pc : mdiobus_free+0x44/0x50
lr : devm_mdiobus_free+0x10/0x20
Call trace:
mdiobus_free+0x44/0x50
devm_mdiobus_free+0x10/0x20
devres_release_all+0xa0/0x100
__device_release_driver+0x190/0x220
device_release_driver_internal+0xac/0xb0
device_links_unbind_consumers+0xd4/0x100
__device_release_driver+0x4c/0x220
device_release_driver_internal+0xac/0xb0
device_links_unbind_consumers+0xd4/0x100
__device_release_driver+0x94/0x220
device_release_driver+0x28/0x40
bus_remove_device+0x118/0x124
device_del+0x174/0x420
fsl_mc_device_remove+0x24/0x40
__fsl_mc_device_remove+0xc/0x20
device_for_each_child+0x58/0xa0
dprc_remove+0x90/0xb0
fsl_mc_driver_remove+0x20/0x5c
__device_release_driver+0x21c/0x220
device_release_driver+0x28/0x40
bus_remove_device+0x118/0x124
device_del+0x174/0x420
fsl_mc_bus_remove+0x80/0x100
fsl_mc_bus_shutdown+0xc/0x1c
platform_shutdown+0x20/0x30
device_shutdown+0x154/0x330
kernel_power_off+0x34/0x6c
__do_sys_reboot+0x15c/0x250
__arm64_sys_reboot+0x20/0x30
invoke_syscall.constprop.0+0x4c/0xe0
do_el0_svc+0x4c/0x150
el0_svc+0x24/0xb0
el0t_64_sync_handler+0xa8/0xb0
el0t_64_sync+0x178/0x17c
So the same treatment must be applied to all DSA switch drivers, which
is: either use devres for both the mdiobus allocation and registration,
or don't use devres at all.
The Marvell driver already has a good structure for mdiobus removal, so
just plug in mdiobus_free and get rid of devres.
Mahesh Bandewar [Mon, 7 Feb 2022 22:29:01 +0000 (14:29 -0800)]
bonding: pair enable_port with slave_arr_updates
When 803.2ad mode enables a participating port, it should update
the slave-array. I have observed that the member links are participating
and are part of the active aggregator while the traffic is egressing via
only one member link (in a case where two links are participating). Via
kprobes I discovered that slave-arr has only one link added while
the other participating link wasn't part of the slave-arr.
I couldn't see what caused that situation but the simple code-walk
through provided me hints that the enable_port wasn't always associated
with the slave-array update.
projects / linux.git / log